January started with a run of product updates and ended with the biggest AI market shock so far.
DeepSeek R1
On January 20, the Chinese lab DeepSeek released R1, an open-weights reasoning model that competes with OpenAI's o1 on math and coding. DeepSeek says V3, the base model behind R1, was trained on 2,048 NVIDIA H800 GPUs in about two months. That is a fraction of what US labs are believed to spend.
On Monday the market reacted. NVIDIA lost about $589 billion in market value in a single day.¹
My read: cheaper training is good news for anyone building on top of models. If strong models get cheaper to make, they get cheaper to use, and more use cases start to make sense.
Stargate
The day after R1 came out, OpenAI, SoftBank, Oracle, and MGX announced Stargate, a new company that plans to invest up to $500 billion over four years in AI data centers in the US, starting with $100 billion.² Earlier in the month, Microsoft said it expects to spend about $80 billion on AI-capable data centers this fiscal year.³
The two stories pull in opposite directions. One says frontier models may get cheaper to build. The other bets hundreds of billions that demand for compute keeps growing. I think both can be true. Cheaper models mean more usage, and more usage needs more capacity.
Agents Start Using Computers
On January 23, OpenAI launched Operator as a research preview for Pro users in the US.⁴ Operator has its own browser. It reads the page from screenshots, then clicks, types, and scrolls to finish tasks like booking a table or filling out a form.
It is slow and needs supervision. It also works on sites that have no API, which is most of the web. That is the part worth watching.
Assistants in the Car
On January 13, Google Cloud announced its Automotive AI Agent, built on Gemini. Mercedes-Benz is the first carmaker to use it, in the MBUX Virtual Assistant coming to the new CLA later this year.⁵ It handles multi-turn conversations and remembers context during a drive, so you can ask a follow-up about a restaurant without starting over.
Generative Worlds
If 2023 was generative images and 2024 was generative video, 2025 could be the year of generative worlds. Google DeepMind's Genie 2 turns a single image into a playable 3D environment.⁶ Fei-Fei Li's World Labs, which raised $230 million last year, is building what it calls large world models.⁷ It is early, but gaming, training simulations, and digital twins are obvious first uses.
What This Means for Product Managers
- Plan for model costs to fall. Features that don't pencil out today may next quarter.
- Build AI into existing workflows. A chat window on every screen is not a strategy.
- Enterprise wants ROI. The experimentation phase is ending. AI features need to show measurable value.
- Narrow beats general. Well-defined use cases outperform general-purpose assistants in enterprise settings.
What AI developments are you watching most closely this year?